Enhancing heart disease prediction with stacked ensemble and MCDM-based ranking: an optimized RST-ML approach.

Abstract

Introduction: Cardiovascular disease (CVD) is a leading global cause of death, necessitating the development of accurate diagnostic models. This study presents an Optimized Rough Set Theory-Machine Learning (RST-ML) framework that integrates Multi-Criteria Decision-Making (MCDM) for effective heart disease (HD) prediction. By utilizing RST for feature selection, the framework minimizes dimensionality while retaining essential information.

Methods: The framework employs RST to select relevant features, followed by the integration of nine ML classifiers into five stacked ensemble models through correlation analysis to enhance predictive accuracy and reduce overfitting. The Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) ranks the models, with weights assigned using the Mean Rank Error Correction (MEREC) method. Hyperparameter tuning for the top model, Stack-4, was conducted using GridSearchCV, identifying XGBoost (XG) as the most effective classifier. To assess scalability and generalization, the framework was evaluated using additional datasets, including chronic kidney disease (CKD), obesity levels, and breast cancer. Explainable AI (XAI) techniques were also applied to clarify feature importance and decision-making processes.

Results: Stack-4 emerged as the highest-performing model, with XGBoost achieving the best predictive accuracy. The application of XAI techniques provided insights into the model's decision-making, highlighting key features influencing predictions.

Discussion: The findings demonstrate the effectiveness of the RST-ML framework in improving HD prediction accuracy. The successful application to diverse datasets indicates strong scalability and generalization potential, making the framework a robust and scalable solution for timely diagnosis across various health conditions.